AMDMIGraphX

AMDMIGraphX copied to clipboard

Remove layernorm fusion

6

pfultz2

Preload tiles into LDS to improve performance of pointwise transposes

6

Fixes #3172.

pfultz2

Add optional fp16 rmsnorm conversion pass to fix fp16 accuracy

10

Fixes https://github.com/ROCm/AMDMIGraphX/issues/2556 RMSNorm is used in LLMs like Llama2. Currently the fp16 version can overflow during normalization. This change try to addess it byconvert normalization to fp32.

attila-dusnoki-htec

SWDEV-451864 - Remove the hard coded path in MIGRAPHX_CXX_COMPILER

5

amdclang++/amdclang can be invoked without providing absolute path, since update-alternatives command has been run for these binaries This will also help to remove the hard coded paths in cmake target...

raramakr

Parser changes to handle MatMulIntegerToFloat

4

Changes to MatMul parser to handle the Microsoft Contrib operator MatMulintegarToFloat Since we have the scale and zero points in our operands we can just perform a multiplied after int8...

TedThemistokleous

Needed to be commented out by #3340 - [ ] Uncomment and investigate - [ ] Fix code that causes this if able - [ ] determine if test stops...

TedThemistokleous

Disabled tidy check in #3340 - [ ] Uncomment and investigate - [ ] Fix code if able - [ ] reneable check

TedThemistokleous

Investigate and cleanup tidy errors that arose via #3340 with bugprone-empty-catch - [ ] Uncomment this in CmakeLists.txt - [ ] determine if false positive or requires fix - [...

TedThemistokleous

Enable attention by default on mi300

3

pfultz2

causten